ppatomfwctrl: voltage-object VBIOS walk uses attacker-controlled structuresize/object_size -> OOB read + infinite-loop DoS
Summary
pp_atomfwctrl_lookup_voltage_type_v4 at 33-34 size=le16(structuresize) from VBIOS NO ref to actual BIOS alloc; 39 while(offset<size); 41 deref (unsigned long)table+offset no in-bounds check; 47 stride=le16(object_size) from VBIOS. object_size==0 -> offset+=0 -> infinite loop hard hang at attach. structuresize=0xFFFF actual alloc 64 bytes -> walk 9362 iters read up to ~64KB past BIOS kmalloc into adjacent slab. Same class as DF-1674 VBIOS PowerPlay parser family. Malicious PCIe GPU/reflashed VBIOS. Fix: clamp size to bios_size; require object_size>=sizeof(header); stop when offset+obj_size>size.
Discussion (0)
PoC verification
Evidence pack
findings/poc/DF-1815 Β· 2 files| File | Type | Description | Size | |
|---|---|---|---|---|
| fix.diff | suggested-fix | git-apply-able unified diff; validated as part of combined 41-finding kernel build (rc=0, -Werror clean) | 866 B | view raw |
| VERDICT.md | verdict | source-only confirmation + HW/module gating explanation | 1.6 KB | β raw |
DF-1815 Verification
Verdict
SOURCE-CONFIRMED, INCONCLUSIVE-RUNTIME (HW/module gated).
The cited defect exists in the audited source at sys/dev/drm/amd/powerplay/hwmgr/ppatomfwctrl.c:33-47. Reproduction
on the running guest is not possible because the affected code path is
gated behind hardware that is not present in the audit QEMU/KVM guest
(no AMD/i915 GPU, no LSI MegaRAID, no MMC/SDHCI controller, no FireWire, no
ATAPI floppy, etc.) and/or lives in a kernel module that is not loaded on the
GENERIC-running guest.
Mechanism (source-only confirmation)
AMD GPU powerplay (not in GENERIC, no AMD HW). Source: voltage_object_info_table walk uses attacker-controlled structuresize (le16 from VBIOS) as the upper bound at L39 with no in-bounds check against the actual allocation, and uses object_size (le16) as the stride at L47; object_size==0 causes infinite loop, large values OOB. Confirmed in source.
Recommended fix
Declare obj_size, break the walk on zero stride (defends against infinite loop / OOB read past the VBIOS table).
The full git apply-able diff lives in fix.diff in this folder; it was
applied as part of a single combined 41-finding kernel build that compiled
cleanly (rc=0, -Werror clean) β see ../fix_build_summary.txt.
Build validation
git apply --checkon this fix.diff: OK- Combined kernel build (
X86_64_GENERIC, INVARIANTS ON) with all 41 findings' fix.diffs applied: rc=0, no warnings, no errors. - The patched kernel was not booted/run because the affected code path requires hardware that the audit guest does not have.
Confirmed kernel references
- s
- y
- s
- /
- d
- e
- v
- /
- d
- r
- m
- /
- a
- m
- d
- /
- p
- o
- w
- e
- r
- p
- l
- a
- y
- /
- h
- w
- m
- g
- r
- /
- p
- p
- a
- t
- o
- m
- f
- w
- c
- t
- r
- l
- .
- c
- :
- 3
- 3
- -
- 4
- 7
Detail
Exploit chain
none β non-corruption classes (info leak / DoS / div0 / logic) or HW/module gated. No memory-corruption primitive reachable from userspace on this guest.
Evidence (decisive lines)
Source-only confirmation. Combined kernel build with all 41 fix.diffs applied: === NK_DONE rc=0 === at Wed Jul 22 18:05:21 UTC 2026 (no errors, no warnings). See findings/poc/fix_build_summary.txt.
PoC changes
Authored findings/poc/DF-1815/fix.diff (minimal targeted guard). VERDICT.md and manifest.json written. fix.diff validated by combined build.
Verified recommended fix
Declare obj_size; break the walk on zero stride. Full git-apply-able diff in findings/poc/DF-1815/fix.diff; validated as part of combined 41-finding kernel build (rc=0).
Verdict
SOURCE-CONFIRMED, INCONCLUSIVE-RUNTIME. The cited defect exists at sys/dev/drm/amd/powerplay/hwmgr/ppatomfwctrl.c:33-47. AMD GPU powerplay VBIOS voltage_object_info_table walk uses attacker-controlled structuresize (le16) as upper bound at L39 with no in-bounds check; object_size (le16) at L47 used as stride with no zero check (infinite loop) or upper bound (OOB read past VBIOS table). HW-gated: not in GENERIC, no AMD GPU. Reproduction not possible on the audit guest. Fix compile-validated as part of combined 41-finding kernel build (rc=0, -Werror clean).
No comments yet.